Sparks Will Fly: engineering creative script conflicts

ABSTRACT

Scripts are often dismissed as the stuff of good movies and bad politics. They codify cultural experience so rigidly that they remove our freedom of choice and become the very antithesis of creativity. Yet, mental scripts have an important role to play in our understanding of creative behaviour, since a deliberate departure from an established script can produce results that are simultaneously novel and familiar, especially when others stick to the conventional script. Indeed, creative opportunities often arise at the overlapping boundaries of two scripts that antagonistically compete to mentally organise the same situation. This work explores the computational integration of competing scripts to generate creative friction in short texts that are surprising but meaningful. Our exploration considers conventional macro-scripts – ordered sequences of actions – and the less obvious micro-scripts that operate at even the lowest levels of language. For the former, we generate plots that squeeze two scripts into a single mini-narrative; for the latter, we generate ironic descriptions that use conflicting scripts to highlight the speaker's pragmatic insincerity. We show experimentally that verbal irony requires both kinds of scripts – macro and micro – to work together to reliably generate creative sparks from a speaker's subversive intent.

KEYWORDS:

• Scripts
• blends
• Twitter
• Twitterbots
• metaphor
• narrative

1. Running with the blend

Whenever we aim to capture the drama and the comedy of the human condition on the stage or on the screen, we first set out to find the right script. So it is hardly surprising that when scholars set out to model our cognitive faculties for understanding all this comedy and drama, they also look to the notion of a script, not as it is written on the page but as it is abstracted in the mind. Because these mental scripts (in the sense of Raskin, 1985; Schank & Abelson, 1977 ) capture the regularities of life and our experiences of the world, they allow us to explain our past and to predict our future. We thus call upon scripts to guide our behaviour whenever we order coffee, make dinner, catch a train or go on a date. At the cognitive level, a script is the codification of a rational expectation, and so the more diverse the set of scripts that a shared culture (bolstered by personal experience) can put at our disposal, the more sophisticated the expectations we can rationally infer. Conversely, a greater diversity of scripts also creates the potential for greater ambiguity: when a situation appears to conform to multiple scripts that codify very different expectations, which should be triggered and which should remain inactive?

The best examples of scripts come, unsurprisingly, from the stage and the screen. Yet even in a theatrical setting the notion of a script can be applied a multiple levels of granularity and specificity. We naturally think of the script as providing a highly specific sequence of stage directions and speech acts for the actors to faithfully follow. But writers themselves follow a more general set of story-telling principles that we might also consider meta-scripts, inasmuch as they codify salient expectations about expectations. As Anton Chekhov famously put it, “ If you say in the first chapter that there is a rifle hanging on the wall, in the second or third chapter it absolutely must go off. If it's not going to be fired, it shouldn't be hanging there ” Bill (1987, p. 79). Objects become physical signifiers for the mental scripts that govern their usage, thus triggering those scripts in the minds of an audience. Though Chekhov also spoke of these scripts as implicit promises to the audience, noting “ It's wrong to make promises you don't mean to keep”, a creative writer is free to depart from the script conventionally associated with a familiar object or idea if the departure can be sufficiently motivated by reference to the original script.

Comedians often construct imaginary situations where departure from the script is the only rational course of action. To make sense of a joke we are forced to re-examine the conventional script that has led us to a logical impasse, to perhaps conclude that the wrong script, and the wrong set of promises, was accidentally triggered (Raskin, 1985) . Or it may be that a creative thinker simply takes issue with the cultural foundations of a script, and seeks redress by exposing the cracks in those foundations. For example, when the serial divorcée Zsa Zsa Gabor was informed by a male interviewer that she was not the “ Good Housekeeping” stereotype of a good wife, she responded: “ Darlink, actually I am an excellent housekeeper. Whenever I leave a man, I keep the house” . Scripts, though apparently rigid constructs, are subjected to constant renegotiation by creative uses such as these. But Gabor's example, like Chekhov's, shows us that scripts attach themselves to loaded words as well as to loaded weapons. They even attach themselves to morphemes such as “un” (is “an untieable knot” a knot that can be untied or a knot that cannot be tied in the first place?) and “kind” (as in e.g. cummings' coinage “manunkind”). Scripts likewise govern our use of hashtags on Twitter, our use of scare-quotes in texts and our use of rabbit-earred quote-unquote hand gestures in spoken conversation. But to be governed is not to be enslaved, and – as our experiments here will show – creative thinkers remain free to renegotiate these conventions and bend scripts to their own meanings.

Every script departure entails a script arrival, for to depart from a conventional script requires that we find another script – perhaps one of our own creation, as in Gabor's retort above – to follow instead. This second script, if chosen with care, can highlight serious deficiencies in the conventional view. Consider this short extract from the movie Jurassic Park, which captures an exchange between the park's creator, John Hammond, and a wry mathematician, Ian Malcolm, who has been asked to evaluate the park's viability before it is opened to the public. The park of the title is populated with genetically engineered dinosaurs, and so the dialogue takes place against a backdrop of carnivorous mayhem:

Table

John Hammond:	All major theme parks have delays. When they opened Disneyland in 1956, nothing worked!
Dr. Ian Malcolm:	Yeah, but, John, if The Pirates of the Caribbean breaks down, the pirates don't eat the tourists.

At this point in the movie, nothing is working in Jurassic Park, but nothing worked in 1956 at Disneyland either, and the latter turned out to be a huge financial and cultural success. Hammond thus frames Disneyland as a script by focusing on the temporal sequence of events associated with its launch, its initial problems, and its eventual success. With this implicit analogy to Jurassic Park, whose launch has been plagued by unique problems of its own, Hammond predicts that his own troubled venture will follow the same script and achieve the same success. In effect, he sees Disneyland and Jurassic Park as two overlapping frames or scripts, and wants others to see the overlap too, so they might come to the same conclusions. Malcolm's rejoinder is also intended to be understood in the context of this analogy, but it is much more than an analogy. It involves mapping, yes, so that The Pirates of the Caribbean is aligned with the attractions of Jurassic Park and the pirates of the former are mapped to the dinosaurs of the latter. But the salient behaviours of the latter – such as eating others willy-nilly – are also integrated with the protagonists of the former, to generate a counterfactual image of animatronic pirates eating tourists in mouse-earred caps. In the words of Fauconnier and Turner (2002), Malcolm has created a blend and is now running it: that is, he is conducting a mental simulation to explore the emergent possibilities that were hitherto just latent in the juxtaposition of both frames or scripts.

Disneyland and Jurassic Park are very different in so many ways, but one gets no sense of these differences from Hammond's analogy, which is designed to emphasise the convergence of scripts and to downplay their divergence. Malcolm's rejoinder, in contrast, takes this convergence as given (hence his “Yeah”) but gives most emphasis to the divergence (hence his “but”). Malcolm's remarks make sport of what Raskin (1985), Attardo and Raskin (1991) and Attardo, Hempelmann, and Maio (2002) call the underlying script opposition (SO), yet he goes further than simply pointing out the SO: he blends both scripts into a ridiculous mélange that forces Hammond (and us) to see the dramatic consequences of the SO. That Malcolm's remark is funnier than Hammond's can be attributed to this use of script overlap and opposition (where Hammond's has just the former), but it is his use of blending that transforms the SO into a ridiculous situation worth laughing at. This transformational effect suggests that blending accounts of frame-shifting – e.g. by Coulson (2001) – offer more than merely notational variants of the script switching accounts of Raskin (1985) and Attardo et al. (2002).

Script blending concretises the SO at the heart of a joke in a way that a simple switch from one script to another cannot. In this paper we model and evaluate these processes of SO and blending not by reference to human-crafted cinematic examples but to the outputs of an autonomous computational system. We explore SO and blending at both the macro-level, the level of narrative, actions and events, and the micro-level, the level at which a meaning is packaged with the appropriate markers to achieve the desired communicative effect. In the course of the paper we shall gradually shift our focus from conventional high-level scripts, conceived as ordered sequences of dramatic actions, to low-level scripts, conceived as simple textual elements that can likewise create rational expectations in the minds of an audience. We show how these expectations can be cultivated and subverted for creative effect by an automated system. The paper culminates with a consideration of one such creative effect, irony, where we evaluate the combined contribution of macro-level and micro-level script elements in the effective communication of an ironic world-view. Few aspects of human creativity are as subtle or as difficult to formalise as irony, but our experiments will show that a computer can exploit tacit scripting conventions to reliably communicate a sense of irony that humans can appreciate. More importantly, by imbuing a generative system with a recognisable capacity for ironic expression, we demonstrate that we can now treat cognitive scripts as more than a convenient theoretical abstraction, and actually use and combine scripts robustly in a practical, large-scale creative generation system.

2. Related work and ideas

Scripts assume a protean form in modern approaches to humor. From Raskin's (1985) use of the classic artificial intelligence (AI) scripts of Schank and Abelson (1977) – in essence, temporally ordered narrative skeletons with roles and variables – scripts have, since Attardo et al. (2002), evolved into generic graph structures. This generalisation turns scripts into structures not unlike the mental spaces of blending theory used by Coulson (2001) and Fauconnier and Turner (2002).

The general theory of verbal humour (or GTVH) of Attardo and Raskin (1991) and Attardo et al. (2002) offers a script-based incongruity-resolution view of verbal humour, in which a narrative may seem compatible with multiple scripts, one of which will be more contextually primed than others. Like its progenitor, Raskin (1985)'s semantic script theory of humour (or SSTH), and its successor, the ontological semantic theory of humour (or OSTH) of Raskin, Hempelman, and Taylor (2009), the GTVH suggests that humour occurs when a listener is deceived into applying a strongly primed script that ultimately leads to a semantic or pragmatic incongruity (see, e.g. Ritchie, 1999; Veale, 2004 ), an impasse of interpretation that can only be resolved by switching to an apparently less salient script. The GTVH views humorous resolution, whether partial or complete, as the work of a particular logical mechanism (LM) that applies at the level of scripts. Understandably, LMs have proved to be the most important elements of the GTVH, prompting Attardo et al. (2002) to enumerate a taxonomy of 27 different LMs.

For the most part, the GTVH/SSTH/OSTH notion of script is grounded in the work of Schank and Abelson (1977), who view scripts as schematic structures that impose a sequential, causal ordering on a narrative and which reflect a single top-down interpretation of events based on an abstracted distillation of relevant episodic memories. This Schankian approach has proven most useful for narrative understanding (see, e.g. Dyer, 1983 ) but it has also been applied to narrative generation (e.g. Meehan, 1981). GTVH scripts can be activated in one of three ways: lexically (by association with a single word, called the lexical handle of the script); sententially (by a pattern of words and lexical scripts); and inferentially, as a by-product of common-sense reasoning (e.g. as when one intuits that a joke is racist and activates a Racism script). Furthermore, since certain elements in a script will be more salient and foregrounded than others, these elements are marked to distinguish them from less salient background elements. Attardo et al. (2002) augmented this general view with a graph-theoretic account of script representation that views scripts as arbitrarily complex symbolic structures, to which mathematical processes such as sub-graph isomorphism can be applied, while more recently, Raskin et al. (2009) have provided an ontological representation (in the computational rather than philosophical sense of ontology) to formalise the makeup of these symbolic structures. This representational shift allows the GTVH/OSTH to encompass even punning as a script-level operation, provided the notion of script is sufficiently generalised to accommodate phonetic as well as semantic and pragmatic expectations.

In the original SSTH of Raskin (1985), the incongruities that trigger script switching are purely semantic, based on a fixed catalogue of antonymies such as DEATH/LIFE. In contrast, the GTVH and OSTH offer the possibility that incongruities may arise from a more nuanced spectrum of conceptual and pragmatic concerns. In a further evolution of the SSTH, the GTVH additionally sees humour production as the culmination of a modular process involving a variety of complementary knowledge resources, each determining a different element of a joke, such as narrative structure, LM, word choice and so on. The LM is the most crucial, but also the most controversial, modular element of the GTVH. For instance, Attardo et al. (2002) suggest that an LM named false-analogy is central to those jokes whose humor derives from ill-judged comparisons, as in the old joke where a mad scientist builds a rocket to the sun but plans to embark at night to avoid being cremated. Here a false-analogy is created between the sun and a light-bulb, suggesting that when the sun is not shining it is not “turned on”, and hence, not hot. In the same vein as Chekhov's broad injunction to playwrights, we might consider false-analogy and its logical kin as meta-scripts whose purpose is to support the humorous conjoining of more specific scripts, such as a script for turning on/off a light-bulb or a script for travelling into space. Hofstadter and Gabora (1994) describe these meta-scripts as joke skeletons, since they provide a common logical structure to whole families of related jokes. As Hofstadter has suggested, one might think of a meta-script as the abstracted remains of a once-elaborate “Ur” -joke whose flesh is now lost, but whose skeleton continues to serve a role in the generation of new creative artifacts.

We can consider our cognitive faculty for metaphor to be another such LM for the juxtaposition and integration of rational expectations from two different domains of experience. For metaphors involve a juxtaposition of the schematic representations of two domains, a source and a target (see Lakoff & Johnson, 1980), and so a striking (and perhaps humorous) metaphor can be viewed as a case of script overlap between two conflicting scripts whose SO gives us the metaphor's characteristic semantic tension. Veale (2014) uses the cut-up technique to generate novel metaphors via a splicing of propositions from very different domains, propositions that are chosen to maximise rather than to minimise the boundary friction between each domain. Such metaphors can themselves be viewed as scripts, allowing a computational system to generate high-friction juxtapositions of conflicting metaphors for the very same topic. Veale, Valitutti, and Li (2015) demonstrate how the @MetaphorMagnet Twitterbot of Veale (2014) harvests metaphorical schemas – such as history is a line and history is a chain – from Web corpora, and treats them as conflicting scripts to generate provocative tweets such as “@war_poet says history is a straight line but @war_prisoner says it is a coiled chain”.

Here the template “ X says T is $S_1$ but Y says it is $S_2$” serves as a meta-script, the textual instantiation of a LM for juxtaposing two conflicting conceptual metaphors. The conflict is anchored in the antonymy of coiled and straight, but it could just as well pivot on an affective opposition between a strongly positive $S_1$ and a strongly negative $S_2$ (or vice versa). As described in Veale (2015) , @MetaphorMagnet also invents its own aptly named interlocutors (such as @war_poet and @war_prisoner ) as imaginary agents to embody the conflicting positions and thus wrap an additional layer of social conflict around the underlying SO. But linguistic and rhetorical style can also be viewed as a script, inasmuch as it brings with it a rich set of norms and expectations. Veale (2015) takes its cue from Raymond Queneau's famous Exercises in style, an Oulipo-inspired exploration of the role of textual style in shaping meaning in varied ways to evoke alternating responses in the reader, from pathos to detachment to laughter (Queneau, 1981; Oulipo, 1983). By giving @MetaphorMagnet a diversity of voices with which to frame its metaphorical outputs – ranging from the philosophical to the religious to the cynical to the childish – Veale (2015) shows how humorous conflict can often be created between the conceptual level of a tweet and its linguistic rendering.

Twitter offers fertile ground to humans and machines alike for this kind of stylistic blend, while the concision required of tweets ensures that such blends are often short, concentrated doses of verbal ingenuity. The comedian Patton Oswalt has initiated a sequence of tweets with the tag #JamesEllroyStarWars that elicits the best examples of this kind of blend, by encouraging followers to blend plot points from the film Star Wars with the distinctive, argot-laden and free-flowing style of American crime writer James Ellroy (noted for L.A. Confidential). Some typical tweets include “Obi-Wan was preaching that Jedi beatnik bebop while an imperial cruiser counted the hairs on our backsides” and “Leia kissed Luke on the mouth. Deep down she knew he was her brother, but she grooved on it”. Another trending tag that encourages stylistic blends is #ThingsJesusNeverSaid. This tag elicits pseudo-religious aphorisms with a big dollop of irony, such as “Love your enemy unless it makes you uncomfortable” and “Blessed are the corporations, for they shall be called my constituents”. So stylistic blends must have something interesting to say, while parodying the most identifiable verbal mannerisms of a well-known communicator. In this vein, @MetaphorMagnet puts the essence of authorial voices such as Yoda, the Hulk, Donald Trump and Jesus Christ into script form, and for the content of its tweets – the second script in its blend – it uses the outputs of Veale (2014)'s Flux Capacitor, to which we turn next. As a taster, this tweet from @MetaphorMagnet aims to capture the essence of Mr. Trump:

Table

Tweet 1:	Blessed are the broke tramps that get jobs, for they shall inherit empires and become wealthy kings.
	#ThingsTrumpNeverSaid #Tramp =#King

In the spirit of Attardo et al. (2002) we thus adopt a rather flexible notion of a script in this work, one that is much less rigid than that of either Schank and Abelson (1977) or Raskin (1985). As used here, a script is any distillation of a conventional viewpoint, in a form that is either conceptual or linguistic, or perhaps both at once if the intent is to capture both style and substance in the same structure. A meta-script – or what Hofstadter and Gabora (1994) call a joke skeleton – is likewise any lexico-conceptual structure that allows us, or our computational systems, to juxtapose two other scripts for creative effect. These lower-level scripts will exhibit a certain mutual antagonism by virtue of their lexical affect – so that positive terms are juxtaposed against negative terms – or by virtue of semantic opposition. While we allow that ambitious script conflicts can pivot around a clash of world-views that exhibit neither of these lexico-semantic markers, we do not focus on such conflicts here, preferring instead to consider those juxtapositions that can be computationally modelled with the symbolic resources harvested in Hao and Veale (2010b), Veale and Li (2011) and Veale (2014).

3. Scripts we live by

Scripts can vary widely in their temporal scale and event resolution. We can thus use scripts to model almost anything, from the actions of sub-atomic particles to the growth of the universe and most everything in between. We often conceive of scripts as bundles of linked actions for everyday events such as using an ATM or ordering lunch (see Dyer, 1983; Schank & Abelson, 1977), but scripts can be just as useful in organising our understanding of events that can last a lifetime. A glance at the obituaries page in any newspaper reveals our desire to impose a linear narrative on a person's life, allowing us to appreciate the life less ordinary as a departure from the scripted norm. From an AI perspective, it can make sense to model people – or rather, categories of people – as scripts, so as to understand their actions as either normative (via script adherence) or transgressive (via script violation).

To generate its condensed stories of change, Veale (2014)'s Flux Capacitor system explicitly models diverse person types, or what Veale calls stereotypes, as scripts. At its core, the system rests on a large set of knowledge triples that characterise diverse kinds of person, such as criminal, surgeon or clown, via their various actions, settings and goals. The Flux Capacitor tags these triples with integers to impose upon them a partial ordering, to indicate, for instance, that surgeons must enroll in medical school (step 0) before they can study medicine (step 1) and graduate with a medical degree (step 2). Any triple may be tagged in this way with an integer from 0 to 9, where 0 indicates a category-entry action (e.g. enrolling in medical school, joining the circus); 9 indicates a category-departure action (e.g. losing one's medical license, getting fired from the circus); 5 indicates steps that one associates with a category instance in full flight (operating on patients as a qualified surgeon or performing pratfalls as a circus clown); 1, …, 4 indicates an action leading up to this mid-life high; and 6, …, 8 marks those actions that take a person a step closer to a final category exit (such as losing one's sense of humour or being sued for malpractice). In the course of one's scripted life, a person is expected to progress from 0 to 9, passing through a series of intermediate steps that draw one fully into a category before inexorably pushing one out again. To generate a normative plot, the Flux Capacitor need only sample its set of triples for a given stereotype to form a chain of successive actions linking any step 0 to a step 9.

But the life less ordinary does not progress from 0 to 9 within the same category. Rather, just as interesting people deviate from the norm, interesting characters deviate from the script by jumping the rails from one category into another. Ebenezer Scrooge goes from a “ grasping, scraping, clutching, covetous old sinner” to secret Santa in the space of A Christmas Carol, jumping categories from misanthropist to philanthropist to find himself following a very different script. Judah Ben-Hur goes from Jewish prince to Roman slave to Arab horseman to ace charioteer in the space of his eponymous epic, while Maximus Decimus Meridius follows a similar, script-hopping trajectory from Roman general to Spanish slave to Roman hero in the film Gladiator.

Entertaining plot twists turn hunters into prey, underdogs into champions, friends into enemies, sinners into saints, or members of one category into something else entirely. To maximise the potential for script conflict, the Flux Capacitor shunts a character into categories and scripts that dramatically flip at least one of its key qualities, such as strong$\to$weak or rich$\to$poor, as when a Tramp inherits the crown to become a King, or when a tech geek becomes a revered tech guru in this example from @MetaphorMagnet:

Table

Tweet 2a:	How might an unpopular geek become a venerated founder?
Tweet 2b:	What if unpopular geeks were to acquire social graces, start businesses and become venerated founders? #Geek= #Founder

@MetaphorMagnet, as a user of the Flux Capacitor, chooses Geek and Founder to serve as the start and end categories for the trajectory of its unnamed character. It reasons that since geeks are typically unpopular and lacking in social graces, this trajectory is likely to yield an interesting script conflict, since founders are typically venerated. As argued in Veale (2012), these stereotypical qualities may be considered part of the scripts for Geek and Founder since they motivate the actions of members of those categories. The scripts for both are stitched together by linking an exit event for Geek (acquiring social graces, a step 9 action) to an entry event for Founder (starting a business, at step 0). In the next section we set about motivating this change of script with an inciting event.

4. Slow-burn transformations

Franz Kafka's Metamorphosis presents a famous literary transformation that is very much the exception rather than the norm. Gregor Samsa, his doomed protagonist, goes from beloved son to hated vermin in an off-stage transformation before the story even begins. This radical change remains unexplained to the end, in what is itself a meta-plot point. But most dramatic changes in characterisation are motivated by observable plot actions. For example, Walter White, the crystal meth-cooking anti-hero at the heart of the TV drama Breaking Bad, goes from caring chemistry teacher to ruthless drug lord when he learns that he has terminal cancer and no longer has anything to lose. To motivate the movement of a character from one stereotypical category (and script track) to another, a generative system must introduce an inciting event that spurs the character to switch tracks. This incitement may come from another character that exercises a strong influence on our protagonist, or it may arise gradually, in the manner of a character's core actions. Let us look at the latter case first, using these @MetaphorMagnet tweets as examples:

Table

Tweet 3a	Tweet 3b	Tweet 3c
#Monotony is when you:	#Gore is when you:	#Cruelty is when you:
1. Invent a fiction	1. Go on a trip	Go on an exploration
2. Become a storyteller	2. Become a holiday maker	2. Become a searcher
3. Tell monotonous stories	3. Enjoy bloody trips	3. Conduct cruel searches
4. Turn into a drudge	4. Turn into a killer	4. Turn into a sadist
#Storyteller =#Drudge	#Holiday_maker =#Killer	#Searcher =#Sadist

The skeletal structure of these stories is necessitated by Twitter's 140-character limit, yet a bare bones rendering is sufficient for each to unfussily tell its tale. In each story, our unnamed protagonist hops from category to category and script to script, opening with a step 0 action from a source script A that permits entry into the category of A. The next stage in each plot is now a blended action that moves the actions of the source category of A into a target category and script of B, where A and B exhibit an obvious contrast. For example, A and B may have opposing lexical affect, as in Storyteller (a mildly positive lexical concept) versus Drudge (a strongly negative lexical concept), or A and B may represent opposite ends of a power relation, as in Master versus Slave or Lord versus Peasant.

Though it is in the nature of story-tellers to tell stories and it falls on drudges to perform monotonous tasks, a blend of both of these behaviours can be observed in the action tell monotonous stories. For if one tells enough stories monotonously, the act of story-telling becomes a chore for speaker and listener alike, causing our protagonist to slowly become a drudge. Two pieces of information are key to arriving at this insight: the first, that drudges typically do monotonous things, is to be found in the script for Drudge. The second, that stories can sometimes be monotonous to tell, is a domain-bridging possibility that is found in neither script but only acquired via experience of the world. In @MetaphorMagnet's case, the outside world is experienced vicariously through the Google Web n-grams (Brants & Franz, 2006), which suggest to the system, via the bigram “monotonous stories” (frequency = 41), that telling stories can sometimes be like performing chores. In the words of Fauconnier and Turner (2002), @MetaphorMagnet recruits the idea that stories may be monotonous as a means of blending together two competing frames.

The stories above squeeze as much as they can into Twitter's minimal containers, framing their four-act plots with a moralising label such as “ #Cruelty is when you ______”. But this additional real-estate can also be used to add an additional fifth act to our plots:

Table

Tweet 4a	Tweet 4b
1. Apply for a scholarship	1. Enroll in a university
2. Become a student	2. Become a student
3. Undergo bizarre educations	3. Undergo brutal exams
4. Turn into a freak	4. Turn into a beast
5. Lose your scholarship	5. Flunk out of university

These five-act tweets leave no room for adornment with hashtags of the form #A=#B, though the meanings remain clear in each case. Each mini-narrative shares the same skeletal structure: 1. an entry action for category A; 2. a statement announcing the protagonist's arrival in category A; 3. a blended action transposing a central event in category/script A into the domain of category/script B; 4. a statement announcing the protagonist's arrival in category B; 5. a subsequent departure from category A. Note how each tweet pairs a category-entry action with a category exit action that governs the same very object (i.e. scholarship in tweet 4a and university in tweet 4b). This minimises the complexity of each narrative and also yields a satisfying symmetry, yet this symmetry can itself be subverted to introduce more ambiguity and nuance. If we switch the fifth act in these tweets, ending tweet 4a with Flunk out of university and ending tweet 4b with Lose your scholarship, the reader would still infer when reading tweet 4a that the protagonist's grant application was successful, though ultimately for naught, and would just as likely infer in tweet 4b that the protagonist had earned a scholarship that was later squandered through freakish abandon. Nor is the system forced to conclude either tweet negatively, for it can just as easily choose the exit action Graduate from university for the fifth act of each. Though this might yield an incongruous ending, it is an incongruity that readers can easily resolve by inferring that brutish abnormalities are no barrier to academic success; indeed, as our protagonist undergoes bizarre educations, it may instead prove to be an advantage. In short, the system has a wealth of combinations to explore.

Though Twitter gives @MetaphorMagnet very little room to maneouvre, it is possible to slow the transformation of a protagonist from an instance of A to an instance of B. The following tweets do not assume an exit from category A, but a blend of A and B:

Table

Tweet 5a	Tweet 5b	Tweet 5c
#Reformer =#Loser	#Mother =#Cynic	#Artist =#Sinner
1. Fight for reform	1. Give birth to a daughter	1. Develop an aesthetic
2. Become a reformer	2. Become a mother	2. Become an artist
3. Launch failed crusades	3. Nurture bitter sons	3. Create illicit pictures
4. Become a failed reformer	4. Become a bitter mother	4. Become an illicit artist
5. Get called a “loser”	5. Get called a “criminal”	5. Get called a “sinner”

So our protagonist remains in category A throughout, yet executes the actions of a member of this category in the manner of someone from category B. The transition between acts 3 and 4 is left deliberately vague in each mini-narrative, to hint at further off-stage actions that eventually lead others to brand the protagonist a member of “B”. @MetaphorMagnet adds the scare quotes to instill doubt and create an ironic distance between that which is quoted and that which is actually meant. In the next section we shall attempt to quantify just how much distance these scare-quotes can instill in the right context.

5. Generating friction with an ironic frame of mind

At first blush, irony can seem an unnecessarily roundabout, complex and even risky mode of communication. The ironic speaker pretends to think $f\left(X\right)$ about a topic X, but in fact thinks $\mathrm{not}\left(f\right(X\left)\right)$, and artfully modulates the expression of the pretend meaning $f\left(X\right)$ to allow an intelligent listener to unpack the intended attitude $\mathrm{not}\left(f\right(X\left)\right)$. If it seems that the speaker has missed a trick by not expressing $\mathrm{not}\left(f\right(X\left)\right)$ more directly, this is to miss the point of irony, which is far from reveling in indirectness for its own sake. Though an ironic speaker runs the risk of being misunderstood and associated with a potentially unpopular proposition $f\left(X\right)$, this speaker actually satisfies multiple communicative goals at once: firstly, that $f\left(X\right)$ is a rational expectation in the current context; secondly, that this expectation has, for some reason, been thwarted; and thirdly, that this failure of rational expectations is deserving of some criticism. Since irony depends crucially on a shared understanding of what those expectations are and on how they have been dashed, it is a phenomenon whose verbal expression depends just as much on a mastery of scripts.

The successful communication of an ironic stance requires an artful combination of scripts, often at multiple levels of granularity. For instance, to evoke the expectation whose failure is the focus of ironic criticism, a speaker will strive to evoke the corresponding script in the minds of an audience. So when one is speaking of what it means to be a “gentleman”, one will rely on the word itself to carry the associated expectations of gentility, civility and chivalry, but an additional script may be explicitly triggered to allow the statement to focus on a particular aspect of gentlemanly behaviour, such as one's expected treatment of women. To show how this expectation has been thwarted, the speaker may evoke a conflicting script to portray the opposite of gentlemanly behaviour, and juxtapose these two scripts so that the latter undermines the former. This juxtaposition may employ what Attardo et al. (2002) call the LM of false-analogy, to offer a comparison that is actually an expression of conceptual distance. To further focus the ironic effect, the speaker may additionally employ any number of micro-scripts. For instance, scare-quotes might be used to highlight a very specific failure of expectations with regards to a single quality or category, while – on Twitter at least – a speaker may also use the hashtag #Irony to convey a generally ironic disposition. Consider the following combination of scripts, both macro- and micro-level, in a @MetaphorMagnet tweet that aims to convey an ironic stance toward a specific topic, modern “gentlemen”:

Table

Tweet 6:

#Irony: When some ladies are pursued by “cultured” gentlemen the way prey is pursued by feral predators. #Lady =#Prey #Gentleman =#Predator

Tweets such as these are constructed by @MetaphorMagnet using two extensive symbolic resources, on the order of tens of thousands of semantic triples apiece: the stereotypical property model of Hao and Veale (2010a), which has been harvested from “ X is as P as Y” similes on the web, and the stereotypical action model of Veale and Li (2011), which has been harvested from the questions that web-users pose to search engines such as Google (e.g. “ why do dogs bury bones?”).

Labels and their conditions of use often become separated by convention, so that the label “ gentleman” is applied to men who are anything but gentle. We can think of our idealised model of a gentleman as a script that provides both stereotypical qualities, such as chivalrous, well-mannered and cultured, and stereotypical actions, such as the pursuit and wooing of ladies. We can likewise think of our idealised model of a predator as a script that provides qualities such as feral and ruthless and actions such as the pursuit and consumption of prey. As argued in Hao and Veale (2010b) , such scripts provide rational expectations of behaviour that can be honored in stand-alone instances and wittily subverted by creative juxtapositions. Our scripts for Gentleman and Predator overlap in a least one schematic action – instances of each pursue instances of a certain other kind of entity – and conflict in at least one specific quality: gentlemen should be cultured, not feral, while feral predators are typically anything but cultured. The shared action, $pursue(X,Y)$, allows a speaker to align the scripts in a putative analogy, while the conflicting qualities, if sufficiently highlighted, mark this as a false-analogy.

In the terminology of Raskin (1985), the script overlap is the action pursue while the semantic opposition (or SO) is the clash of cultured and feral. So what marks this analogy as a false analogy is the conflict between cultured (a word with a perceptible positive affect) and feral (a word with a perceptible negative affect). To make sense of the analogy one of these qualities must give way to the other: are predators more cultured that we conventionally think, or are gentlemen now more feral than they once were? The phrasing of the analogy suggests the latter, but the speaker reinforces this interpretation by placing scare-quotes around “cultured”. When used in this way, quotation marks serve as an appeal to an authority from which the speaker seeks to establish some ironic distance. The quoting of “cultured” here thus suggests that the word is not used to convey the speaker's own view, but to ironically echo the default (but now sadly outdated) viewpoint of others (Sperber & Wilson, 1981). The scare-quotes thus serve as a micro-script whose function is to direct additional scrutiny to the fault-line of the juxtaposion.

The explicit #Irony is much less surgical in its action, and may even – to use the logic of (Grice, 1967, 1989) – seem an affront to common-sense in its overt signalling of a phenomenon that is prized for its subtlety. Yet the marking of irony is commonplace on Twitter, where speakers often mark content that might otherwise be misconstrued and attract not plaudits but a barrage of criticism. In fact, Attardo (2000) argues that irony is often marked in spoken language too, albeit never so explicitly, by our choice of words and syntax, while Reyes, Rosso, and Veale (2013) show that a wide variety of lexical and syntactic features can be exploited by a machine-learning system to reliably (if far from perfectly) recognise ironic intent in human-crafted tweets. The use of an explicit hashtag is just one more micro-level script in the arsenal of an ironic speaker, and one might well pose the question as to how effectively this script, and indeed others, contribute to the successful communication of an ironic stance. In the above example we see all scripts working together to construct an ironic pretense (that gentlemen are still “cultured” even when they act as feral predators) that is designed to be penetrated by a knowing audience (Clark & Gerrig, 1984; Wilson, 2006). But how much does each script contribute to the ironic effect? If we can quantify this contribution, we can better understand how script-based reasoning, at the macro- and micro- levels, contributes to linguistic creativity more generally.

To quantify each script's contribution to an ironic stance we must first operationalise what we mean by “irony”. In the above example we expect an intelligent reader to infer that gentlemen who act like predators are not cultured after all, even if they retain a veneer of culture in their actions. A word carries a positive lexical affect to the extent that speakers are inclined to view it as a pleasant or complimentary description of an entity, while a word carries a negative lexical affect to the extent that speakers view it as an unpleasant or perjorative label. Lexical affect is a subtle dimension of nouns such as “rain” and “wind” but a more obvious dimension of qualities such as “rainy”, “boring” and “unsophisticated”. We can expect the lexical affect of a word like “cultured” to be highly positive in a null context, to reflect its default meaning, but to be significantly diminished in a context that holds this aspect of the Gentleman stereotype up to ridicule. Moreover, we should be able to quantify the extent of this downshift in lexical affect to show how irony undercuts the received wisdom of the stereotype. Thus, if we measure the mean lexical affect of the quality cultured (or whatever positive quality is focalised by the ironic statement) in a null context, we can expect a substantially positive score on a scale running from −1.0 to +1.0. If we later measure the mean lexical affect of the focalised word in a critical utterance where it is undermined by irony, we should observe a significant downward shift in its mean positivity (see Gardiner & Dras, 2012, who operationalise the valence shifting phenomenon, and Simančík & Lee, 2009, who used valence shifting to improve the performance of sentiment analysis). The shift may be so dramatic that its affect in context flips from positive to negative, so that “cultured” becomes a signifier of scorn.

We have tested this hypothesis using multiple variants of 80 randomly sampled ironic tweets from @MetaphorMagnet's outputs, to determine which combinations of scripts are most effective at achieving a substantial downshift in affect for the focalised word. Human judges on the crowd-sourcing platform CrowdFlower were asked to rate the lexical affect of the focalised word in each of the 80 tweets, but in a null context. That is, the judges were presented with each word in isolation, and asked to score its default positivity on a scale from $-1.0$ to +1.0. The mean rating for each quality gives us a baseline against which to judge the effect of ironic downshifting in the context of a specific utterance. Judges were then presented with structural variants of the 80 test utterances, to tease out the relative contribution of different scripts. All of the sampled tweets have the same linguistic structure even if they speak to, and criticise, different stereotypical beliefs, so we can illustrate these variant conditions using our example “cultured” tweet:

Table

base:	Some ladies are pursued by cultured gentlemen.
quot:	Some ladies are pursued by “ cultured” gentlemen.
comp:	Some ladies are pursued by cultured gentlemen the way prey is pursued by feral predators.
quot+comp:	Some ladies are pursued by “ cultured” gentlemen the way prey is pursued by feral predators.
hash+comp:	#Irony: When some ladies are pursued by cultured gentlemen the way prey is pursued by feral predators.
hash+quot+comp:	#Irony: When some ladies are pursued by “ cultured” gentlemen the way prey is pursued by feral predators.

The above variants were presented to different judges on CrowdFlower, and 10 judgements of the lexical affect of the focal word were elicited for each context. The judgements of scammers and other disengaged users were automatically removed by peppering the test stimuli with “ gold-set” questions and answers. Table 1 shows the mean lexical affect of the focal word on the scale −1.0 to +1.0 when judges are presented with different structural variants. Lexical affect is a continuous dimension that we score on a gradient, while the positivity rate is simply the rate at which judges consider the focal word to be a positive description (which is, to say, more positive than negative) in context.

Table 1. Mean lexical affect and positivity rate, with standard deviations.

Variant	Lexical affect	Positivity rate
BASE	0.51 ± 0.38	0.91 ± 0.15
QUOT	0.41 ± 0.46	0.82 ± 0.13
COMP	0.29 ± 0.49	0.75 ± 0.15
QUOT   +   COMP	0.20 ± 0.54	0.64 ± 0.16

We can observe a general progression at work here: scare-quotes reduce the perceived positivity of the focal word, but not as much as the analogical construction comp, yet both conditions – macro-script and micro-script – work well together (quot   +   comp) to reduce the perceived positivity even further and strengthen the ironic effect. Table 2 shows the mean downshift in positivity for each structural variant.

Table 2. Mean differences in affective downshift across alternate experimental settings.

Variant 1/Variant 2	QUOT	COMP	QUOT + COMP
BASE	−0.10	−0.22	−0.31
QUOT	–	−0.12	−0.21
COMP	–	–	−0.09

We can now see the relative return for an ironic speaker in using a particular variant, or combination of variants, to communicate a failure of expectations and achieve a critical downshift in positivity for the word that denotes the greatest expectation. The juxtaposition of opposing scripts (comp) yields the highest individual return, but this can be improved with the addition of scare-quotes (quot) to achieve an even stronger effect. So let us now consider the most overt structural variant of all, the use of the hashtag #Irony. One would expect that if ironic downshifting of positivity is our goal, signalling this goal to the reader with an overt request for an ironic reading will add to the magnitude of the shift even if it diminishes the wit and subtlety of the utterance. We take as our baseline the comp variant, since this achieves the greatest individual downshift, and consider whether the addition of the hash condition – alone, or in combination with the quot condition – can achieve a significant decrease in perceived affect.

Surprisingly, as shown in Table 3, we see that the explicit marker #Irony carries relatively little weight on Twitter, where intuitions about what is and is not irony are far from convergent. After Bonferroni correction is applied to the initial significance value, the addition of the overt hashtag #Irony to the comp juxtaposition of scripts does not appear to produce a truly significant downshift in affect, and contributes even less when it is combined with scare-quotes (quot   +   hash). Irony is a diffuse concept, and on social media the word can denote anything from a sardonic attitude to a humorous coincidence. So to tell a reader that an utterance is ironic appears to be doubly detrimental as a communicative strategy. Not only does the admission rob the utterance of its appearance of subtlety and wit, it does not aid significantly in the search for the author's intended meaning. This, perhaps, is the most reliably ironic quality we can associate with the hashtag #Irony.

Table 3. Mean downward shift in lexical affect when combining the COMP baseline with either QUOT, HASH or QUOT   +   HASH.

Baseline variant	QUOT	HASH	QUOT + HASH
COMP	−0.21	−0.11	−0.22
	p<.001	p<.03	p<.001

Note: The reported p-values have undergone Bonferroni correction.

Because so many of the human users that will interact with our creative language systems will be fluent producers and consumers of irony, the reliable communication of an ironic stance becomes a necessary aspect of any system that seeks to interact with humans on their own terms. As such, irony generation is an intrinsic task of no little social value. Our experimental framework applies a novel twist to the study of irony by using test stimuli that are wholly machine-generated. Since the underlying computational system (MetaphorMagnet) is capable of generating hundreds of thousands of novel figurative utterances on demand, where each meets the same exacting specification, we believe we have demonstrated the important role that computational creativity can play in the study of human creativity. Any researchers who wish to exploit these machine-crafted utterances in their own work have three ways of accessing this data. First, the Twitterbot @MetaphorMagnet generates new utterances on a hourly basis, and its timeline already contains over 20,000 timestamped tweets. Second, the MetaphorMagnet web-service provides individual metaphors and analyses on demand, for user-specified topics. Third, researchers can obtain bulk data – on the order of thousands or hundreds of thousands of instances of specific tweet-types – simply by containing the authors directly. In this way we encourage others to build on the fruits of automated linguistic creativity, and to replicate the experimental findings reported in this paper and in our related work.

Nonetheless, our interest in irony here is chiefly as an extrinsic task, insofar as the delicacy of irony presents us with an ideal test-case for quantifying the relative effect of different kinds of script to creative communication. Our experiments support the established view in humour research (formalised by Raskin, 1985 ) that creative wit arises from a conflict of scripts, the conflict of which is itself engineered by a higher level meta-script-like structure called a LM (Attardo & Raskin, 1991). But our analysis also demonstrates a significant role for low-key elements of communication, such as scare-quotes, which can themselves be seen as a kind of micro-script. These micro-scripts operate at the surface level of language, and mean very little in themselves, but in combination with higher level scripts they become valuable shapers of meaning and potent elements of creative communication. To understand how humans use conventions to shape and/or subvert expectations, so that our machines might creatively do likewise, we must understand how scripted conventions of all kinds and at all levels of granularity can contribute to the construction of a creative stance.

6. Conclusions

The blending of stereotypical categories and scripts is not so very different from William Burroughs’ famed use of the cut-up method to generate random and potentially creative texts (see Burroughs, 1963; Chamberlain & Etter, 1983; Veale, 2014). Except, of course, that a system such as @MetaphorMagnet must serve both as scissors and critic, to create cut-ups from its knowledge-base of scripts and their possible renderings in natural language (whether as chunks of dialogue or as declarative sentences), and it must also evaluate and filter the outputs, to keep the few that actually work and to reject the multitude that do not. What may seem at first to be the most pressing constraint on @MetaphorMagnet's choices – Twitter's size limit on tweets – turns out to be the least vexing; indeed, it transpires that there are many ways of using this constraint to a system's advantage. For Twitter users expect tweets to suggest much more than could ever be explicitly said in 140 characters, and eagerly bridge the chasm between what is expressly stated and what is presumably intended. Users who must engage their imaginations as a matter of course on Twitter do not feel short-changed by a system that asks them to put flesh on the barest of narrative bones.

We have seen here how an author, whether natural or artificial, can shape the affect perceived by a reader in ways that range from the subtle to the overt. The goal here is not to gain the reader's agreement that “yes, this is ironic” but – as in the Valentino system of Guerini, Strapparava, and Stock (2008) – to influence the reader's perception of the meaning and affect of the ideas in an utterance. That, after all, is the communicative purpose of irony: though some set out to construct an ironic pose, most speakers use irony to convey a failure of expectations and to guide a reader toward a shared feeling of disappointment with events and the world. To measure the effectiveness of a system that strives for the creative juxtaposition of familiar ideas, we do not set out to judge the labels assigned to its outputs by a human evaluator, but to quantify the extent to which these outputs actually shift a human judge's preconceived view of those ideas.

George Orwell once claimed that every joke is a tiny revolution because jokes disrupt: they highlight the hidden fault lines in our received wisdom and show how sensible conventions can, in certain combinations, yield entertaining surprises. More generally, creative language of all kinds can be a disruptive force in how we speak and how we think. For good metaphors, like good jokes, disrupt. Though conventional metaphors ultimately become part of the wallpaper of language, so much so that we hardly notice them any more, even these can conceal deep ruptures in the underlying structure of our conceptual systems. The comp variant in the previous section makes these ruptures apparent in the clash of scripts and expectations, even if it takes the quot variant to direct our gaze to the point of maximal disruption. For metaphors become wittily creative when they make a virtue of the ruptures that separate two ideas and their associated mental frames or scripts. As Pollio (1996, p. 248) put it, “split reference yields humour if the joined items (or the act joining them) emphasise the boundary or line separating them; split reference yields metaphor if the boundary between the joined items (or the act joining them) is obliterated and the two items fuse to form a single entity”. By dialing up the antagonism between different domains, or between reality and fantasy, or between expectation and observation, a juxtaposition of scripts can yield a witty, eye-catching and thought-provoking utterance that is worthy of sharing on a social platform such as Twitter. If the overt hash micro-script asks a reader to “ let there be light”, only the comp and quot scripts actually create the sparks than can ignite the reader's imagination.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Alessandro Valitutti  http://orcid.org/00000-0002-1523-5967.

Additional information

Funding

This research was supported by the EC project WHIM: The What-If Machine (www.whim-project.eu). This work was also supported by Seventh Framework Programme [611560].